singhmandavi/math-slm-qwen2.5-0.5b-v3

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The singhmandavi/math-slm-qwen2.5-0.5b-v3 is a 0.5 billion parameter Qwen2.5-0.5B-Instruct based small language model, fine-tuned by team03 for step-by-step mathematical problem-solving. It specializes in GATE-style and general mathematics, providing detailed solutions for complex math questions. With a context length of 32768 tokens, this model is optimized for competitive exam aspirants and math students requiring free-form math QA.

Loading preview...

Model Overview

This model, singhmandavi/math-slm-qwen2.5-0.5b-v3, is a specialized small language model (SLM) developed by team03 for the Pramana SLM++ Bootcamp. It is built upon the Qwen2.5-0.5B-Instruct base model, featuring 0.5 billion parameters and fine-tuned using QLoRA (r=16, 4-bit NF4 base, bf16 compute) to excel in mathematical problem-solving.

Key Capabilities

  • Step-by-step Math Problem Solving: Designed to provide detailed, chain-of-thought reasoning for complex mathematical questions, including GATE-style problems.
  • Domain-Specific Focus: Optimized for general mathematics and competitive exam preparation.
  • Instruction Following: While generally capable, it exhibits a verbose CoT bias, sometimes prepending preambles even under strict output constraints.

Training and Performance

The model was trained on a corpus of 29,853 rows, including both direct answers and Chain-of-Thought (CoT) examples, sourced from ExamBench and MathNet IMO split. Training was conducted for 3 epochs with a max sequence length of 2,048 tokens. Evaluation shows healthy training metrics, though a marginal convergence failure was noted. The model's instruction-following score is 24.24%, with a primary limitation being its tendency to prepend CoT preambles.

Good For

  • Competitive Exam Aspirants: Particularly those preparing for GATE or similar mathematics-focused exams.
  • Math Students: Requiring detailed, step-by-step solutions to mathematical problems.
  • Free-form Math QA: Handling queries that benefit from explicit reasoning processes rather than just direct answers.

Finetunes
1 models